A Mobile Terminal-based Nomogram for Early Predicting Severity of Acute Pancreatitis

Background: Early prediction of the severity of acute pancreatitis (AP) is important but there is no preferred method in China. We aimed to develop and validate a simple-to-use predictive nomogram for persistent organ failure (POF) on admission in patients with AP. Methods: Data from 816 consecutive patients was obtained from internal (Chengdu) retrospective datasets and formed the training cohort for nomogram development. Data from 398 and 880 consecutive patients from internal (Chengdu) and external (Nanchang) prospective datasets formed the validation cohorts (all admitted < 48 hours of symptom onset). Univariate and multivariate logistic regressions were used to identify independent prognostic factors to establish the nomogram for POF. The calibration curves, concordance index (C-index), decision curve analysis (DCA), and clinical impact curve (CIC) were used to evaluate the performance of the nomogram and its clinical utility. The area under the receiver-operating characteristic curve (AUC) with 95% CI and likelihood ratio as well as post-test probability were applied. Measurements and main results: Age, respiratory rate, albumin, lactate dehydrogenase, oxygen support, and pleural effusion were identied as independent prognostic factors for POF and were included in the nomogram model (web-based calculator: https://shina.shinyapps.io/DynNomapp/). This predictive nomogram had good predictive ability for POF (C-indexes of 0.88, 0.91 and 0.81 for the training and two validation cohorts) and promising clinical utility (DCA: better or equivalent than prognostic scores; CIC: high clinical net benet). The AUC of (0.91 [0.88-0.94] and 0.81 [0.79-0.84]), negative likelihood ratio (NLR 0.11 and 0.29), post-test probability of negative (0.9% and 6.7%) of the nomogram were superior in predicting POF than all other routinely used clinical prognostic scoring systems in both validation cohorts. Similar ndings were observed for predicting major infection (superior to other prognostic scores) and mortality (superior or equally to others). Conclusions: The validated nomogram comprises 6 independent prognostic factors to predict major clinical outcomes of patients with AP in two distinct Chinese centers. This mobile terminal-based nomogram should be validated in other settings and considered for clinical practice and trial allocation, until more accurate biomarkers are discovered. low density lipoprotein-cholesterol; PCD, percutaneous catheter drainage; CTSI, computed tomography severity index. in AP the early of Our results showed a positive association between of POF respiratory rate or requirement for oxygen support on both respiratory rate and oxygen for their convenience and prognostic value. The signicance of pleural effusion in AP patients has been long reported [70], and is part of BISAP [40]. Our results showed that pleural effusion (odds ratio: 3.61 95%CI 1.97–6.6, P < 0.001) an independent predictor for SAP, consistent with study


Introduction
Acute pancreatitis (AP) is a protean and heterogenous disease with a spectrum of severity ranging from mild to critical [1]. The early prediction of the severity of AP is a cornerstone of management because it informs clinical decisions about triage, transfer and intervention [2]. Early prediction is also important in the research setting, where the accurate allocation of patients into trial arms based on predicted severity is critical for the testing of treatments for AP. The key determinants of AP severity are organ failure and infected pancreatic necrosis [3,4]. With the recent improvements in treating infected pancreatic necrosis, persistent organ failure (POF) has become the most important determinant for mortality [5][6][7][8][9], and is the basis for the grade of severe AP (SAP) in the revised Atlanta classi cation (RAC) [10]).
Since the Ranson score was introduced in 1974 [11], more than 20 prognostic scores have been studied for AP severity prediction [12]. However, their clinical utility is limited by an accuracy of predicting POF at circa 75% and many of them are cumbersome to use [13]. Recent systematic reviews and meta-analyses conclude that current early predictors of POF [14], infected pancreatic necrosis [14], and mortality [12] do not have su cient accuracy for the decision making in individual patients. The ideal predictor of POF would be applied on patient admission and within 24 hours of the onset of symptoms. It would be cost-effective, easy to use and have an accuracy between 95-100%. Considerable progress has been made in identifying serum biomarkers to stratify early risk and severity in patients with AP [15]. However, due to their rapid time-course changes in serum concentration, nonspeci city, cost, complexity, and suboptimal accuracy, none of the biochemical biomarkers have been adopted into routine clinical practice. In recent years, different nomograms have been applied for predicting severity (mortality) [16][17][18][19], splanchnic vein thrombosis [20], abdominal infection [21], computed tomography index for assessing AP outcomes [22] and catheter drainage in necrotizing AP [23,24] as well as for oral refeeding intolerance during hospital stay [25] and new-onset diabetes after AP [26] (Table 1). However, of the 4 studies investigated the predictive value of nomograms for severity or mortality, 3 [17]) and 1 was retrospective [19] in nature. Besides, none of these studies reported the time of symptom onset to hospital admission. Therefore, the early prediction of AP severity at primary admission or early transfer remains to a challenge.

Study design and ethics
This study followed the STROBE guidelines [27] for observational studies. The study protocol was approved by respective Institutional Review Board in these two hospitals. Data were obtained from two AP cohorts in West China Hospital of Sichuan University (WCH/SCU): retrospective datasets between 1st July 2009 and 30th June 2013 as the training cohort [9]; prospective datasets between 1st January 2016 and 31th August 2017 as the internal validation cohort [28]. For the purpose of external validation, we obtained the datasets from AP database of The First A liated Hospital of Nanchang University from January 2005 to December 2012 [29]. Exclusion criteria included patients admitted to another hospital prior to WCU/SCU, re-admitted during the same episode of AP, chronic pancreatitis, pancreatic neoplasia, trauma, or pregnancy as AP etiologies, and advanced pre-existing comorbidities consistent with previous studies [6, 9,28,30].

Inclusion and exclusion criteria
The external validation cohort also contained transferred patients and the cases with the nal formulized nomogram factors missing were excluded.

De nitions, variables, and outcome measures
Demographic features collected on admission, such as age, gender, underlying disease to score Charlson co-morbidity index and abdominal pain onset time to admission. Vital signs, laboratory parameters (routine blood and biochemical indices), details of oxygen treatment and presence of pleural effusion that were mostly obtained from nonenhanced computed tomography scan on admission. Experienced PhD students and resident doctors specializing in management of AP were stringently trained for data collection according to pre-de ned proforma and standard operating procedures in both centers. Each proforma was checked and signed off by attending or more senior doctors.

Etiologies
Hypertriglyceridemia as the etiology was de ned as admission serum triglyceride (TG) level > 1000 mg/dl (11.3 mmol/l) or TG > 500 mg/dl (5.65 mmol/l) with lipemic serum or a previous history of hypertriglyceridemia [31,32]. Biliary etiology was considered if gallstones or biliary sludge was present on radiological imaging or had alanine aminotransferase over three times the upper limit of normal [33,34]. Alcohol excess was referred if with drinking history > 35 standard drinks per week for > 5 years [35] or was deduced according to our center's own practice according to the drinking history after ruling out other etiologies.
Organ failure, local complication, and major infection POF was de ned as at least one of the systems (respiratory, circulatory, and renal) having a SOFA score ≥ 2 and lasting ≥ 48 hours [36]. Acute necrotic collection and acute peripancreatic uid collection were de ned as per RAC criteria [10]. Major infection included infected pancreatic necrosis, bacteremia, or pneumonia alone or in combination [37]. Mortality (including those who were automatically discharged with persistent multiple organ failure and had high possibility of death) and length of hospital stay were recorded for the index hospitalization.
The prognostic scores evaluated in this study were National Early Warning Score (NEWS) [38], Systemic In ammatory Response Syndrome (SIRS) [39], Bedside Index for Severity in Acute Pancreatitis (BISAP) [40], modi ed Glasgow criteria [41], Acute Physiology and Chronic Health Examination (APACHE) II [42], and Sequential Organ Failure Assessment (SOFA) [43]. These were all calculated on admission (or within 6 hours of admission if absent earlier in minority of cases).

Statistical analysis
Continuous variables were expressed as median with 25th -75th percentile and were compared by Mann-Whitney U test (2 groups) or Kruskal-Wallis H test (3 groups). Categorical data were reported as number with percentage and were compared by means of χ 2 or Fisher's exact test.
The signi cance of each variable in the training cohort was assessed by univariate logistic regression analysis for investigating the independent risk factors of POF. All variables signi cantly associated with POF were candidates for stepwise multivariate analysis and the results were used to formulate a nomogram. The predictive performance of the nomogram was measured by concordance index (C-index) and calibration with 1000 bootstrap samples to decrease the over t bias [44]. Decision curve analysis (DCA) and clinical impact curve (CIC) were used to evaluate the clinical utility of the novel model [45]. DCA is a simple, novel method of evaluating predictive models with decision analyses of diagnostic and prognostic tests by using a risk-bene t ratio [46]. CIC is another type of plot produced by Decision Curve that shows the estimated number who would be declared to having high risk for each risk threshold and visually shows the proportion of those who are cases (true positives) [45].
The area under curve (AUC) of the receiver operating characteristic (ROC) curves with 95% con dence intervals (CI) of all potential predictors were calculated. An AUC of 0.5, 0.7 to 0.8, 0.8 to 0.9, and > 0.9 was suggestive of no discrimination, acceptable, excellent and outstanding, respectively [47]. (http://www.Rproject.org). Two-tailed P value with statistical signi cance set at < 0.05 was used for all tests.

Patient characteristics
There were 816 and 398 patients in the training and internal validation cohorts, respectively, that met the eligibility criteria (Additional le 1: Figure S1A, B). After removing missing values, the external validation cohort contained 880 records. The demographic pro les of the three cohorts are outlined in   Table 3). These variables were nally used to construct our predictive nomogram (Fig. 1A). To facilitate the clinical application of our ndings, we developed a mobile terminal-based calculator (https://shina.shinyapps.io/DynNomapp/) of the predictive nomogram (Fig. 1B, C). In the training cohort, the C-index of the nomogram was 0.88 ( Fig.   2A) with good consistency between the predicted SAP and the actual POF observed shown by the calibration curve (Fig. 2B). In the internal validation cohort, the C-index of the nomogram for predicting SAP reached 0.91 (Fig. 2C) and with a better consistency (Fig. 2D) compared to the training cohort.
The C-index was 0.81 (Fig. 2E) in the external validation cohort and the calibration curve also showed satisfactory consistency (Fig. 2F).

Discussion
In this study, we developed and validated a mobile terminal-based nomogram for predicting POF, major infection, and mortality using six independent prognostic factors (age, respiratory rate, albumin, lactate dehydrogenase, oxygenation, and pleural effusion) that readily available on admission. This nomogram was found to be superior with the both internal and external validation cohort compared with other prognostic scores recommended for clinical use for predicting POF and major infection. As POF is the diagnostic criteria for SAP, this nomogram is recommended for routine clinical use to predict SAP on admission or to rule it out (validation NLR 0.11/0.29). Therefore, these ndings encourage the use of the simple-to-use web-based nomogram before new markers are developed and introduced in our settings. The consecutive nature of patient recruitment and short time from onset of pain to admission, stringently applied in two large different Chinese centers, adds strength to these conclusions.
Age is recognized as an individual risk factor for increased severity of AP and has been used by several prognostic scores [12,13] and practice guidelines [49][50][51]. To investigate the role of age and comorbidity in the severity of AP, Frey et al. [52] carried out a retrospective study in 84,713 patients with a rst-attack AP. They found that the 65 to 75 age group, and age > 75 are strong predictors of early death with an odds ratio (OR) of 2.6 and 5.2, respectively. Similar ndings also applied to patients with two chronic comorbidities (OR: 3.5) or ≥ 3 comorbidities (OR: 7.4). Moreover, the mortality rate was only 0.1% (14/14,280) for younger patients (age < 55) without chronic comorbidities compared to 5.9% (701/24,852) for elderly patients (age > 64) with ≥ 3 comorbidities in the rst 14 days. In addition, they showed that recent cancer, heart failure, and renal and liver diseases are strongly correlated with outcomes. Further, in acute interstitial AP, which is known to have low mortality, the Charlson comorbidity index was strongly associated with adverse clinical outcomes [53]. Because of the signi cant impact of the degree and number of comorbidities on clinical outcomes, we therefore excluded patients with advanced (end-stage) comorbidities with an emphasizing on assessing the intrinsic prognostic factors for AP severity.
Hypoalbuminemia occur in critically ill patients due to several factors including dilution from resuscitation, increased interstitial loss, altered liver function, and catabolic nutritional state [54]. It is strongly associated with poor clinical outcomes in acutely ill patients [55] and it has also been shown to independently associated with POF and mortality in AP patients [56,57]. Whitcomb et al.
[58] has recently found that albumin dropped rapidly in AP patients with multiple organ failure resulting in unregulated capillary leak with continued loss of larger plasma proteins. In contrast, the plasma albumin levels only dropped slightly in patients without multiple organ failure who tended to recover quickly unless they develop complications (infected pancreatic necrosis or sepsis). The therapeutic effect of albumin in in ammatory states is not only by affecting plasma volume dilation, but also by regulating in ammation and oxidative stress [59,60]. Therefore, serum albumin level has been incorporated in some AP severity prognostic indices (Glasgow criteria [41] and nomogram [18]).
Raised lactate levels has been observed in many critical acute illness situations including sepsis [61] and AP [62-64]. Elevated lactate may serve as a protective mechanism and has been shown to reduce Toll-like receptor and in ammasome-mediated pancreatic and liver injury via its receptor GPR81 [65]. LDH can reversibly catalyze the oxidation of lactate to pyruvate and has been employed by Ranson, Glasgow, Japanese Severity Score [13], and a nomogram [17] for early AP severity prediction and reported as a simple and useful parameter for predicting POF [66, 67], and pancreatic necrosis [68].
Moreover, urinary LDH has also been reported as an useful biomarker for septic acute kidney injury [69].
Respiratory rate and oxygen support re ect respiratory status and with respiratory failure as one most common organ dysfunction in AP [9], the early recognition of respiratory dysfunction is considered important. Our results showed a positive association between the development of POF and an increased respiratory rate or requirement for oxygen support on admission. Therefore, both respiratory rate and oxygen support have been adopted in NEWS [38] for their convenience and prognostic value. The signi cance of pleural effusion in AP patients has been long reported [70], and is part of BISAP [40]. Our results showed that pleural effusion (odds ratio: 3.61 95%CI 1.97-6.6, P < 0.001) was an independent predictor for SAP, consistent with a previous study [71].
In our univariate analysis before establishing the predictive nomogram, we also found white blood cell count [13], glucose [13,72], urea (or blood urea nitrogen) [6, 13], creatinine [13], and ionized calcium [13] were independent individual prognostic factors for POF, consistent with previously published literature. However, when these parameters were tted into our multiple logistic regression model, they only had negligible impact on the nal nomogram. Unlike most previously published studies included high proportion of SAP patients, we used a training consecutive cohort constituted only up to 10% of SAP patients which may partially explain different weights of individual prognostic factors in varied epidemiology situations. For example, three of the four existing predictive nomograms for AP severity were conducted in the ICU settings which cannot be generalizable for emergence departments or general wards where patients are primarily admitted.
Some of the six independent prognostic factors of our nomogram are included as part of NEWS (respiratory rate and oxygen support) and BISAP (age, respiratory rate, and pleural effusion), the two that we found had justi able good predictive values for POF in both our training and internal validation cohorts. However, the nomogram was simpler than BISAP and more AP-speci c than NEWS to quantitatively predict clinical outcomes in a personalized way. In addition, we validated the nomogram in another Chinese tertiary hospital with the etiological composition different from ours. The results showed that the nomogram with stable clinical applicability in both AP cohorts with hypertriglyceridemia (internal validation) or biliary (external validation) as the main etiology, adding strength to its applicability.
Our study also has some limitations. Firstly, the nomogram model was developed mainly based on the variables that were easy to get in our retrospective sets, but did not include other factors that may in uence the precision of the model. For example, the oxygenation index was not included in our analysis because only paucity data were available. In a most recent study [73], the authors found that oxygenation index had low prognostic power (AUC 55.3%) for acute respiratory distress syndrome. Secondly, the nomogram did not have speci c markers for circulatory and renal failure.
The reasons for this may be attributed to low incidence of circulatory and renal failure of the study population and at the early disease stage respiratory failure commonly precedes other organ failures [5,8]. Thirdly, the lack of international validation may limit the extrapolation and generalizability of the nomogram.

Conclusions
Our nomogram based on six readily available factors accurately predicted POF on admission in patients with AP. This nomogram can be routinely used for early AP severity prediction in our clinical practice if further validated in multiple center studies.